Anthropic tests small AI models across low-cost connected devices

Anthropic demonstrated that small AI models could generate output across low-cost connected devices by reducing the bits used to represent model weights and splitting workloads among devices, Ars Technica reports. Its Gemma4-26B-A4B test required 15 boxes with 4GB of memory each. Qwen3.5-9B configured other language models to run on eight of nine hardware types. The company warns that such adaptability could eventually support worms that spread across unfamiliar devices without human intervention. Its experiment nevertheless exposed significant challenges to building a self-replicating AI malware system; the demonstrated ability to run models does not establish that replication capability.